AI Curriculum Pipeline
Sole designer and engineer
An applied-AI pipeline built into The Notwork that produces complete lessons. It runs nine steps modeled on real curriculum development: blueprint, outline, content, editorialize, accuracy check, media, formatting, audio, and release. Each step builds on the last, runs on the model the educator picks from Anthropic, OpenAI, or Google, and can be refined or regenerated before the lesson publishes straight to the platform.
- Per lesson, was ~1 week
- 10–30 min
- Lessons published
- 55
- Pipeline steps
- 9
- Human review
- Every step
What it's built with
The problem
Curriculum production was slow and inconsistent. A single quality lesson could take a week by hand, and the platform needed a steady stream of multi-modal content for students.
What I built
A nine-step pipeline that walks an educator from rough requirements to a published, multi-modal lesson, with refine and regenerate controls at every step and direct publishing into The Notwork.
Inside the product
What's inside
- Nine steps: blueprint → outline → content → editorialize → accuracy check → media → formatting → audio → release
- Per-step model picker across Claude, GPT, and Gemini
- Per-step refine and regenerate with persistent context
- Editorial scoring with suggested line edits
- Fact checks with confidence scores and needs-review flags
- Image generation with approve or regenerate, then CDN upload
- Text-to-speech audio narration
- Generated practice questions
- Direct publish into The Notwork as a content item
The impact
Cut lesson production from about a week to 10–30 minutes, with higher consistency than the manual process. 55 lessons published so far.
What stood out
- Modeled the pipeline on how curriculum teams actually build lessons, so each step has one job and a clear output.
- Model choice per step across Anthropic, OpenAI, and Google (Claude, GPT, and Gemini), including Gemini image generation for lesson media.
- Human-in-the-loop by design: each step takes the previous output as context, and an educator can refine or regenerate it before it feeds the next.
- An editorial step scores tone and readability and proposes line edits as before-and-after diffs; an accuracy step checks each factual claim with a confidence score and flags what needs subject-matter review.
- Generated images wait for an educator to approve or regenerate them, then upload to the Cloudflare CDN and are written into the lesson.
- Each finished lesson ships with text-to-speech narration and practice questions.